Since the release of advanced generative AI models that could produce realistic artistic renditions, write novel-length stories, and can now even create music, the art world has reacted with a flurry of debate over what is art, what tools can be used to create it, and how AI can enter the industry responsibly – if at all.

The copyright debate is coming from both sides of the AI art controversy – from artists who want AI models to stop training on their work, to users who want to copyright their AI-generated creations.

Both sides bring into question greater themes around art, ownership, and the commodification of creation. Intermixed into every argument for and against each copyright debate are questions surrounding the foundation of generative AI models, the nature of their training, what it means to create a work of ‘art,’ as well as data protection rights.

All of this debate is centred in a trying time for artists as a whole, amid the writers and actors strikes taking place in Hollywood to the copyright lawsuits right here in the UK.

Here, we’ll dive into the two major debates over AI and copyright.

1. Can AI-Generated Art be Copyrighted?

Stephen Thaler, an artist and programme developer based in the United States, created an AI image generator dubbed ‘Creativity Machine’ claiming that it had generated a piece of art ‘on its own accord,’ according to recent court documents read. Thaler then tried to copyright the image under his own name, claiming that, as the creator of Creativity Machine, the AI’s output can be credited under himself.

However, the US Copyright Office refused to issue the copyright, and Thaler appealed the case to a US Court. On Friday, Judge Beryl Howell issued her ruling – that human authorship was a “bedrock” for copyright law, and that no copyright had ever been given to something that was “absent of any guiding human hand”.

In her ruling, she did note that the world was changing as AI was advancing and making its way into the mainstream, acknowledging the fact that this may be the first of many cases, many rulings, and these rulings may change.

What is interesting about the case is that Thaler stated his AI algorithm generated the image without human input, which he later attempted to change in court, saying that he “directed his AI to create the Work” and that the machine is under his control. As these statements were not admissible in court, the judge made the ruling that outputs without any human input could not receive a copyright.

The judge recognises the need, however, for further laws surrounding the amount of human input and direction needed in AI-generated works to warrant authorship and copyright.

The case offers two unusual points that may make it difficult to relate to other AI copyright cases – Thaler created his own AI algorithm to generate the image, but also said he played no role in the image’s generation.

AI image and ‘art’ generators typically function different than this case, and are often not exclusively used by their creators. Users instead write prompts for the AI to then generate an ‘output.’

So, while the case is a landmark in refusing copyright licensing for AI images, Thaler plans to appeal, and the case may not be applicable to other copyright attempts.

The US copyright office said in March they would be open to granting an AI-generated image copyright to its ‘creator’ (i.e., the person that prompted the AI model) on a ‘case-by-case’ basis.

In this case, AI simply lost.

2. Using Copyrighted Art to Train AI

On the other side of the debate is the copyrighted material AI models are trained on. These range from original artworks posted on online forums like the long-running DeviantArt, to famous books.

Artists expressed their outrage that their artwork on DeviantArt would be automatically used to train an AI image generator powered by Stable Diffusion, a AI model. Artists were similarly worried with Adobe said that artwork created using their products – ranging from Illustrator to Photoshop – could be used to train their own AI.

Users can opt-out on both forums, however artists have said the rules should instead be opt-in.

Google discretely updated its terms and conditions to include AI training under its ability to scrape data from the internet, and Meta neglected to disclose what data it used to train its LLaMA 2 model when it was launched in July.

Instead of paying an artist to render an artwork, users can simply write a prompt on an AI model asking for an artwork of [insert idea] in the style of [insert art style, ranging from Baroque to classic Disney to Steampunk], and voila: art (?).

But how does an AI know what Baroque, Disney, or steampunk art looks like? By looking at art rendered by those artists, companies, or in that style.

This has prompted ideas on originality, as well as copyright. Writers have signed an open letter to different AI companies, like OpenAI, Alphabet, and Meta, to stop using their work without consent or compensation.

In early 2023, Getty Images opened a lawsuit against Stability AI, which is the creator of Stable Diffusion, claiming the company had processed around 12 million copyrighted images to train its AI system without consent.

Recommended

• Yes-anding AI: Artificial Intelligence Stars at the Edinburgh Fringe
• Gartner: AI Driving Trends in Data Science and ML
• 10% of UK Teachers Struggle to Tell AI and Student Work Apart

Artists are suing DeviantArt saying that their art had been used to train its AI model without their consent, and that they could potentially lose commissions and money because of the AI system.

There still appears to be global confusion on how to handle this issue, and governments appear to be falling behind as artists are already affected by AI’s use of copyrighted work.

The EU AI Act will require models to disclose if any copyrighted material is used for training purposes, but this still does not provide a clear picture on the rights of artists to opt-out of AI training datasets or to seek compensation for their work being used.

Overall, the issue of copyrights, AI, and art is far from over.